Roadmap

Milestones are ordered by risk, not by pipeline order: the consensus quality model is the scientific claim, so it is validated before any of the throughput work is done.

This page is the summary. ROADMAP.md in the repository root carries the per-item state and is the one kept current as work lands.

milestone	content	state
M0	repo, .mig format, FASTQ IO, read simulator	done
M2	checkout: patterns, trimming, header transfer, UMI statistics, paired input, strand normalisation, multi-core, whitelists, dual-end barcodes	done bar .mig bucket output, i7xi5 and the bit-parallel matcher
M1	assemble: consensus, sub-clustering, quality model, --contig, --fast	done bar R1/R2 overlap merge
M3	refine: error model, barcode correction, cell calling, QC	done bar splitting the fitted floor into RT and first-cycle PCR, which needs two chemistries rather than a better estimator
M4	end-to-end, suggest, subsample, plot, notebooks	done, downstream contract measured
M5	benchmarks, isalgo/umi_data, release	in progress

Explicitly out of scope for v2.0

• Alignment and variant calling. MAGERI’s job. The pipeline ends at consensus FASTQ and hands off to arda, minimap2 or bwa-meme.

• Indels. Illumina indel rates are around \(10^{-6}\) per base and there is no dataset in the benchmark set that would let us verify indel handling. Substitutions only, everywhere.

• Duplex consensus (DCS). v2.0 extracts duplex tags and emits single-strand consensuses. Pairing the two strand families is a later addition, and until it exists no error-suppression claim here is based on duplex data.

• EmptyDrops-style cell rescue. Cell calling is OrdMag plus a knee. Reproducing Cell Ranger’s second pass is its job, not ours.